appendix-e
appendix E Additional resources
When you outgrow this book, these are the references and tools to reach for next:
- CUDA programming guide—https://docs.nvidia.com/cuda/cuda-programming-guide/
- cuBLAS documentation—https://docs.nvidia.com/cuda/cublas/
- cuDNN documentation—https://docs.nvidia.com/deeplearning/cudnn/
- CUTLASS GitHub—https://github.com/NVIDIA/cutlass
- NCCL documentation—https://docs.nvidia.com/deeplearning/nccl/
- OpenMPI documentation—https://www.open-mpi.org/doc/
E.1 Profiling and debugging tools
- Nsight Systems—System-wide performance analysis
- Download: https://developer.nvidia.com/nsight-systems
- Use for application timelines, CPU-GPU coordination, memory transfers, stream analysis
- Nsight Compute (NCU)—Kernel-level profile
- Download https://developer.nvidia.com/nsight-compute
- Use for memory bandwidth, compute utilization, occupancy, warp stalls
- Compute Sanitizer—Memory error detection
- Included with CUDA Toolkit
- Use for out-of-bounds access, uninitialized memory, race conditions
E.1.1 Nsight Systems: Application-level profiling
Nsys helps identify bottlenecks in the overall application flow. It’s particularly useful for the following cases: